#include <stdint.h>
#include <stdlib.hfa>

#include <unistd.h>								// STDERR_FILENO
#include <inttypes.h>
#include "bits/debug.hfa"
#include "bits/locks.hfa"
#include "stats.hfa"
#include "strstream.hfa"

#if !defined(__CFA_NO_STATISTICS__)
	void __init_stats( struct __stats_t * stats ) {
		stats->ready.push.local.attempt = 0;
		stats->ready.push.local.success = 0;
		stats->ready.push.share.attempt = 0;
		stats->ready.push.share.success = 0;
		stats->ready.push.extrn.attempt = 0;
		stats->ready.push.extrn.success = 0;
		stats->ready.pop.local .attempt = 0;
		stats->ready.pop.local .success = 0;
		stats->ready.pop.local .elock   = 0;
		stats->ready.pop.local .eempty  = 0;
		stats->ready.pop.local .espec   = 0;
		stats->ready.pop.help  .attempt = 0;
		stats->ready.pop.help  .success = 0;
		stats->ready.pop.help  .elock   = 0;
		stats->ready.pop.help  .eempty  = 0;
		stats->ready.pop.help  .espec   = 0;
		stats->ready.pop.steal .attempt = 0;
		stats->ready.pop.steal .success = 0;
		stats->ready.pop.steal .elock   = 0;
		stats->ready.pop.steal .eempty  = 0;
		stats->ready.pop.steal .espec   = 0;
		stats->ready.pop.search.attempt = 0;
		stats->ready.pop.search.success = 0;
		stats->ready.pop.search.elock   = 0;
		stats->ready.pop.search.eempty  = 0;
		stats->ready.pop.search.espec   = 0;
		stats->ready.threads.migration = 0;
		stats->ready.threads.threads   = 0;
		stats->ready.sleep.halts   = 0;
		stats->ready.sleep.cancels = 0;
		stats->ready.sleep.wakes   = 0;
		stats->ready.sleep.exits   = 0;

		#if defined(CFA_HAVE_LINUX_IO_URING_H)
			stats->io.alloc.fast        = 0;
			stats->io.alloc.slow        = 0;
			stats->io.alloc.fail        = 0;
			stats->io.alloc.revoke      = 0;
			stats->io.alloc.block       = 0;
			stats->io.submit.fast       = 0;
			stats->io.submit.slow       = 0;
			stats->io.flush.external    = 0;
			stats->io.calls.flush       = 0;
			stats->io.calls.submitted   = 0;
			stats->io.calls.drain       = 0;
			stats->io.calls.completed   = 0;
			stats->io.calls.errors.busy = 0;
			stats->io.poller.sleeps     = 0;
		#endif

		#if defined(CFA_STATS_ARRAY)
			stats->array.values = alloc(CFA_STATS_ARRAY);
			stats->array.cnt = 0;
		#endif
	}

	void __tally_stats( struct __stats_t * cltr, struct __stats_t * proc ) {
		__atomic_fetch_add( &cltr->ready.push.local.attempt, proc->ready.push.local.attempt, __ATOMIC_SEQ_CST ); proc->ready.push.local.attempt = 0;
		__atomic_fetch_add( &cltr->ready.push.local.success, proc->ready.push.local.success, __ATOMIC_SEQ_CST ); proc->ready.push.local.success = 0;
		__atomic_fetch_add( &cltr->ready.push.share.attempt, proc->ready.push.share.attempt, __ATOMIC_SEQ_CST ); proc->ready.push.share.attempt = 0;
		__atomic_fetch_add( &cltr->ready.push.share.success, proc->ready.push.share.success, __ATOMIC_SEQ_CST ); proc->ready.push.share.success = 0;
		__atomic_fetch_add( &cltr->ready.push.extrn.attempt, proc->ready.push.extrn.attempt, __ATOMIC_SEQ_CST ); proc->ready.push.extrn.attempt = 0;
		__atomic_fetch_add( &cltr->ready.push.extrn.success, proc->ready.push.extrn.success, __ATOMIC_SEQ_CST ); proc->ready.push.extrn.success = 0;
		__atomic_fetch_add( &cltr->ready.pop.local .attempt, proc->ready.pop.local .attempt, __ATOMIC_SEQ_CST ); proc->ready.pop.local .attempt = 0;
		__atomic_fetch_add( &cltr->ready.pop.local .success, proc->ready.pop.local .success, __ATOMIC_SEQ_CST ); proc->ready.pop.local .success = 0;
		__atomic_fetch_add( &cltr->ready.pop.local .elock  , proc->ready.pop.local .elock  , __ATOMIC_SEQ_CST ); proc->ready.pop.local .elock   = 0;
		__atomic_fetch_add( &cltr->ready.pop.local .eempty , proc->ready.pop.local .eempty , __ATOMIC_SEQ_CST ); proc->ready.pop.local .eempty  = 0;
		__atomic_fetch_add( &cltr->ready.pop.local .espec  , proc->ready.pop.local .espec  , __ATOMIC_SEQ_CST ); proc->ready.pop.local .espec   = 0;
		__atomic_fetch_add( &cltr->ready.pop.help  .attempt, proc->ready.pop.help  .attempt, __ATOMIC_SEQ_CST ); proc->ready.pop.help  .attempt = 0;
		__atomic_fetch_add( &cltr->ready.pop.help  .success, proc->ready.pop.help  .success, __ATOMIC_SEQ_CST ); proc->ready.pop.help  .success = 0;
		__atomic_fetch_add( &cltr->ready.pop.help  .elock  , proc->ready.pop.help  .elock  , __ATOMIC_SEQ_CST ); proc->ready.pop.help  .elock   = 0;
		__atomic_fetch_add( &cltr->ready.pop.help  .eempty , proc->ready.pop.help  .eempty , __ATOMIC_SEQ_CST ); proc->ready.pop.help  .eempty  = 0;
		__atomic_fetch_add( &cltr->ready.pop.help  .espec  , proc->ready.pop.help  .espec  , __ATOMIC_SEQ_CST ); proc->ready.pop.help  .espec   = 0;
		__atomic_fetch_add( &cltr->ready.pop.steal .attempt, proc->ready.pop.steal .attempt, __ATOMIC_SEQ_CST ); proc->ready.pop.steal .attempt = 0;
		__atomic_fetch_add( &cltr->ready.pop.steal .success, proc->ready.pop.steal .success, __ATOMIC_SEQ_CST ); proc->ready.pop.steal .success = 0;
		__atomic_fetch_add( &cltr->ready.pop.steal .elock  , proc->ready.pop.steal .elock  , __ATOMIC_SEQ_CST ); proc->ready.pop.steal .elock   = 0;
		__atomic_fetch_add( &cltr->ready.pop.steal .eempty , proc->ready.pop.steal .eempty , __ATOMIC_SEQ_CST ); proc->ready.pop.steal .eempty  = 0;
		__atomic_fetch_add( &cltr->ready.pop.steal .espec  , proc->ready.pop.steal .espec  , __ATOMIC_SEQ_CST ); proc->ready.pop.steal .espec   = 0;
		__atomic_fetch_add( &cltr->ready.pop.search.attempt, proc->ready.pop.search.attempt, __ATOMIC_SEQ_CST ); proc->ready.pop.search.attempt = 0;
		__atomic_fetch_add( &cltr->ready.pop.search.success, proc->ready.pop.search.success, __ATOMIC_SEQ_CST ); proc->ready.pop.search.success = 0;
		__atomic_fetch_add( &cltr->ready.pop.search.elock  , proc->ready.pop.search.elock  , __ATOMIC_SEQ_CST ); proc->ready.pop.search.elock   = 0;
		__atomic_fetch_add( &cltr->ready.pop.search.eempty , proc->ready.pop.search.eempty , __ATOMIC_SEQ_CST ); proc->ready.pop.search.eempty  = 0;
		__atomic_fetch_add( &cltr->ready.pop.search.espec  , proc->ready.pop.search.espec  , __ATOMIC_SEQ_CST ); proc->ready.pop.search.espec   = 0;
		__atomic_fetch_add( &cltr->ready.threads.migration , proc->ready.threads.migration , __ATOMIC_SEQ_CST ); proc->ready.threads.migration  = 0;
		__atomic_fetch_add( &cltr->ready.threads.threads   , proc->ready.threads.threads   , __ATOMIC_SEQ_CST ); proc->ready.threads.threads    = 0;
		__atomic_fetch_add( &cltr->ready.sleep.halts       , proc->ready.sleep.halts       , __ATOMIC_SEQ_CST ); proc->ready.sleep.halts        = 0;
		__atomic_fetch_add( &cltr->ready.sleep.cancels     , proc->ready.sleep.cancels     , __ATOMIC_SEQ_CST ); proc->ready.sleep.cancels      = 0;
		__atomic_fetch_add( &cltr->ready.sleep.wakes       , proc->ready.sleep.wakes       , __ATOMIC_SEQ_CST ); proc->ready.sleep.wakes        = 0;
		__atomic_fetch_add( &cltr->ready.sleep.exits       , proc->ready.sleep.exits       , __ATOMIC_SEQ_CST ); proc->ready.sleep.exits        = 0;

		#if defined(CFA_HAVE_LINUX_IO_URING_H)
			__atomic_fetch_add( &cltr->io.alloc.fast       , proc->io.alloc.fast       , __ATOMIC_SEQ_CST ); proc->io.alloc.fast        = 0;
			__atomic_fetch_add( &cltr->io.alloc.slow       , proc->io.alloc.slow       , __ATOMIC_SEQ_CST ); proc->io.alloc.slow        = 0;
			__atomic_fetch_add( &cltr->io.alloc.fail       , proc->io.alloc.fail       , __ATOMIC_SEQ_CST ); proc->io.alloc.fail        = 0;
			__atomic_fetch_add( &cltr->io.alloc.revoke     , proc->io.alloc.revoke     , __ATOMIC_SEQ_CST ); proc->io.alloc.revoke      = 0;
			__atomic_fetch_add( &cltr->io.alloc.block      , proc->io.alloc.block      , __ATOMIC_SEQ_CST ); proc->io.alloc.block       = 0;
			__atomic_fetch_add( &cltr->io.submit.fast      , proc->io.submit.fast      , __ATOMIC_SEQ_CST ); proc->io.submit.fast       = 0;
			__atomic_fetch_add( &cltr->io.submit.slow      , proc->io.submit.slow      , __ATOMIC_SEQ_CST ); proc->io.submit.slow       = 0;
			__atomic_fetch_add( &cltr->io.flush.external   , proc->io.flush.external   , __ATOMIC_SEQ_CST ); proc->io.flush.external    = 0;
			__atomic_fetch_add( &cltr->io.calls.flush      , proc->io.calls.flush      , __ATOMIC_SEQ_CST ); proc->io.calls.flush       = 0;
			__atomic_fetch_add( &cltr->io.calls.submitted  , proc->io.calls.submitted  , __ATOMIC_SEQ_CST ); proc->io.calls.submitted   = 0;
			__atomic_fetch_add( &cltr->io.calls.drain      , proc->io.calls.drain      , __ATOMIC_SEQ_CST ); proc->io.calls.drain       = 0;
			__atomic_fetch_add( &cltr->io.calls.completed  , proc->io.calls.completed  , __ATOMIC_SEQ_CST ); proc->io.calls.completed   = 0;
			__atomic_fetch_add( &cltr->io.calls.errors.busy, proc->io.calls.errors.busy, __ATOMIC_SEQ_CST ); proc->io.calls.errors.busy = 0;
			__atomic_fetch_add( &cltr->io.poller.sleeps    , proc->io.poller.sleeps    , __ATOMIC_SEQ_CST ); proc->io.poller.sleeps     = 0;
		#endif
	}

	#define eng3(X) (ws(3, 3, unit(eng( X ))))

	void __print_stats( struct __stats_t * stats, int flags, const char * type, const char * name, void * id ) with( *stats ) {

		char buf[1024];
		ostrstream sstr = { buf, 1024 };

		if( flags & CFA_STATS_READY_Q ) {

			sstr | "----- " | type | "\"" | name | "\" (" | "" | id | "" | ") - Ready Q Stats -----";

			uint64_t totalR = ready.pop.local.success + ready.pop.help.success + ready.pop.steal.success + ready.pop.search.success;
			uint64_t totalS = ready.push.local.success + ready.push.share.success + ready.push.extrn.success;
			sstr | "- totals   : " | eng3(totalR) | "run," | eng3(totalS) | "schd (" | eng3(ready.push.extrn.success) | "ext," | eng3(ready.threads.migration) | "mig)";

			double push_len = ((double)ready.push.local.attempt + ready.push.share.attempt + ready.push.extrn.attempt) / totalS;
			double sLcl_len = ready.push.local.success ? ((double)ready.push.local.attempt) / ready.push.local.success : 0;
			double sOth_len = ready.push.share.success ? ((double)ready.push.share.attempt) / ready.push.share.success : 0;
			double sExt_len = ready.push.extrn.success ? ((double)ready.push.extrn.attempt) / ready.push.extrn.success : 0;
			sstr | "- push avg : " | ws(3, 3, push_len)
			     | "- l: " | eng3(ready.push.local.attempt) | " (" | ws(3, 3, sLcl_len) | ")"
			     | ", s: " | eng3(ready.push.share.attempt) | " (" | ws(3, 3, sOth_len) | ")"
			     | ", e: " | eng3(ready.push.extrn.attempt) | " (" | ws(3, 3, sExt_len) | ")";

			double rLcl_pc = (100.0 * (double)ready.pop.local .success) / totalR;
			sstr | "- local    : " | eng3(ready.pop.local .success) | "-"| ws(3, 3, rLcl_pc) | '%'
			     | " (" | eng3(ready.pop.local .attempt) | " try," | eng3(ready.pop.local .espec) | " spc," | eng3(ready.pop.local .elock) | " lck," | eng3(ready.pop.local .eempty) | " ept)";
			double rHlp_pc = (100.0 * (double)ready.pop.help  .success) / totalR;
			sstr | "- help     : " | eng3(ready.pop.help  .success) | "-"| ws(3, 3, rHlp_pc) | '%'
			     | " (" | eng3(ready.pop.help  .attempt) | " try," | eng3(ready.pop.help  .espec) | " spc," | eng3(ready.pop.help  .elock) | " lck," | eng3(ready.pop.help  .eempty) | " ept)";
			double rStl_pc = (100.0 * (double)ready.pop.steal .success) / totalR;
			sstr | "- steal    : " | eng3(ready.pop.steal .success) | "-"| ws(3, 3, rStl_pc) | '%'
			     | " (" | eng3(ready.pop.steal .attempt) | " try," | eng3(ready.pop.steal .espec) | " spc," | eng3(ready.pop.steal .elock) | " lck," | eng3(ready.pop.steal .eempty) | " ept)";
			double rSch_pc = (100.0 * (double)ready.pop.search.success) / totalR;
			sstr | "- search   : " | eng3(ready.pop.search.success) | "-"| ws(3, 3, rSch_pc) | '%'
			     | " (" | eng3(ready.pop.search.attempt) | " try," | eng3(ready.pop.search.espec) | " spc," | eng3(ready.pop.search.elock) | " lck," | eng3(ready.pop.search.eempty) | " ept)";

			sstr | "- Idle Slp : " | eng3(ready.sleep.halts) | "halt," | eng3(ready.sleep.cancels) | "cancel," | eng3(ready.sleep.wakes) | "wake," | eng3(ready.sleep.exits) | "exit";
			sstr | nl;
		}

		#if defined(CFA_HAVE_LINUX_IO_URING_H)
			if( flags & CFA_STATS_IO ) {
				sstr | "----- " | type | "\"" | name | "\" (" | "" | id | "" | ") - I/O Stats -----";

				uint64_t total_allocs = io.alloc.fast + io.alloc.slow;
				double avgfasta = (100.0 * (double)io.alloc.fast) / total_allocs;
				sstr | "- total allocations : " | eng3(io.alloc.fast) | "fast," | eng3(io.alloc.slow) | "slow (" | ws(3, 3, avgfasta) | "%)";
				sstr | "-     failures      : " | eng3(io.alloc.fail) | "oom, " | eng3(io.alloc.revoke) | "rvk, " | eng3(io.alloc.block) | "blk";

				uint64_t total_submits = io.submit.fast + io.submit.slow;
				double avgfasts = (100.0 * (double)io.submit.fast) / total_submits;
				sstr | "- total submits     : " | eng3(io.submit.fast) | "fast," | eng3(io.submit.slow) | "slow (" | ws(3, 3, avgfasts) | "%)";
				sstr | "- flush external    : " | eng3(io.flush.external);

				sstr | "- io_uring_enter    : " | eng3(io.calls.flush) | " (" | eng3(io.calls.drain) | ", " | eng3(io.calls.errors.busy) | " EBUSY)";

				double avgsubs = ((double)io.calls.submitted) / io.calls.flush;
				double avgcomp = ((double)io.calls.completed) / io.calls.drain;
				sstr | "-     submits       : " | eng3(io.calls.submitted) | "(" | ws(3, 3, avgsubs) | "/flush)";
				sstr | "-     completes     : " | eng3(io.calls.completed) | "(" | ws(3, 3, avgcomp) | "/drain)";

				sstr | "- poller sleeping   : " | eng3(io.poller.sleeps);
				sstr | nl;
			}
		#endif

		if(flags) write( sstr, stdout );
	}

	#if defined(CFA_STATS_ARRAY)
		extern "C" {
			#include <stdio.h>
			#include <errno.h>
			#include <sys/stat.h>
			#include <fcntl.h>
		}

		void __flush_stat( struct __stats_t * this, const char * name, void * handle) {
			int ret = mkdir(".cfadata", 0755);
			if(ret < 0 && errno != EEXIST) abort("Failed to create directory .cfadata: %d\n", errno);

			char filename[100];
			snprintf(filename, 100, ".cfadata/%s%p.data", name, handle);

			int fd = open(filename, O_WRONLY | O_APPEND | O_CREAT, 0644);
			if(fd < 0) abort("Failed to create file %s: %d\n", filename, errno);

			for(i; this->array.cnt) {
				char line[100];
				size_t n = snprintf(line, 100, "%llu, %lld\n", this->array.values[i].ts, this->array.values[i].value);
				write(fd, line, n);
			}

			this->array.cnt = 0;
			close(fd);
		}

		static __spinlock_t stats_lock;

		void __push_stat( struct __stats_t * this, int64_t value, bool external, const char * name, void * handle ) {
			if(external) lock(stats_lock __cfaabi_dbg_ctx2);

			if( this->array.cnt >= CFA_STATS_ARRAY ) __flush_stat( this, name, handle );

			size_t idx = this->array.cnt;
			this->array.cnt++;

			if(external) unlock(stats_lock);

			this->array.values[idx].ts = rdtscl();
			this->array.values[idx].value = value;
		}
	#endif
#endif